Model-Based Compressive Sensing Richard G. Baraniuk, Volkan Cevher, Marco F. Duarte, Chinmay Hegde Abstract Compressive sensing (CS) is an alternative to Shannon/Nyquist sampling for acquisition of sparse or compressible signals that can be well approximated by just K ≪ N elements from an N -dimensional basis. Instead of taking periodic samples, we measure inner products with M<N random vectors and then recover the signal via a sparsity-seeking optimization or greedy algorithm. The standard CS theory dictates that robust signal recovery is possible from M = O (K log(N/K)) measurements. The goal of this paper is to demonstrate that it is possible to substantially decrease M without sacrificing robustness by leveraging more realistic signal models that go beyond simple sparsity and compressibility by including dependencies between values and locations of the signal coefficients. We introduce a model- based CS theory that parallels the conventional theory and provides concrete guidelines on how to create model-based recovery algorithms with provable performance guarantees. A highlight is the introduction of a new class of model-compressible signals along with a new sufficient condition for robust model- compressible signal recovery that we dub the restricted amplification property (RAmP). The RAmP is the natural counterpart to the restricted isometry property (RIP) of conventional CS. To take practical advantage of the new theory, we integrate two relevant signal models — wavelet trees and block sparsity — into two state-of-the-art CS recovery algorithms and prove that they offer robust recovery from just M = O (K) measurements. Extensive numerical simulations demonstrate the validity and applicability of our new theory and algorithms. Index Terms Compressive sensing, sparsity, signal model, union of subspaces, wavelet tree, block sparsity The authors are listed alphabetically and are with the Department of Electrical and Computer Engineering of Rice University, Houston, Texas. Email: {richb, volkan, duarte, chinmay}@rice.edu; Web: dsp.rice.edu/cs. This work was supported by the grants NSF CCF-0431150 and CCF-0728867, DARPA/ONR N66001-08-1-2065, ONR N00014-07-1-0936 and N00014-08-1- 1112, AFOSR FA9550-07-1-0301, ARO MURI W311NF-07-1-0185, and the Texas Instruments Leadership University Program.
45
Embed
Richard G. Baraniuk, Volkan Cevher, Marco F. Duarte ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Model-Based Compressive Sensing
Richard G. Baraniuk, Volkan Cevher, Marco F. Duarte, Chinmay Hegde
Abstract
Compressive sensing (CS) is an alternative to Shannon/Nyquist sampling for acquisition of sparse or
compressible signals that can be well approximated by just K ≪ N elements from an N -dimensional
basis. Instead of taking periodic samples, we measure inner products with M < N random vectors
and then recover the signal via a sparsity-seeking optimization or greedy algorithm. The standard CS
theory dictates that robust signal recovery is possible from M = O (K log(N/K)) measurements. The
goal of this paper is to demonstrate that it is possible to substantially decrease M without sacrificing
robustness by leveraging more realistic signal models that go beyond simple sparsity and compressibility
by including dependencies between values and locations of the signal coefficients. We introduce a model-
based CS theory that parallels the conventional theory and provides concrete guidelines on how to create
model-based recovery algorithms with provable performance guarantees. A highlight is the introduction
of a new class of model-compressible signals along with a new sufficient condition for robust model-
compressible signal recovery that we dub the restricted amplification property (RAmP). The RAmP is
the natural counterpart to the restricted isometry property (RIP) of conventional CS. To take practical
advantage of the new theory, we integrate two relevant signal models — wavelet trees and block sparsity
— into two state-of-the-art CS recovery algorithms and prove that they offer robust recovery from just
M = O (K) measurements. Extensive numerical simulations demonstrate the validity and applicability
of our new theory and algorithms.
Index Terms
Compressive sensing, sparsity, signal model, union of subspaces, wavelet tree, block sparsity
The authors are listed alphabetically and are with the Department of Electrical and Computer Engineering of Rice University,
Houston, Texas. Email: richb, volkan, duarte, [email protected]; Web: dsp.rice.edu/cs. This work was supported by the
grants NSF CCF-0431150 and CCF-0728867, DARPA/ONR N66001-08-1-2065, ONR N00014-07-1-0936 and N00014-08-1-
1112, AFOSR FA9550-07-1-0301, ARO MURI W311NF-07-1-0185, and the Texas Instruments Leadership University Program.
I. INTRODUCTION
We are in the midst of a digital revolution that is enabling the development and deployment
of new sensors and sensing systems with ever increasing fidelity and resolution. The theoretical
foundation is the Shannon/Nyquist sampling theorem, which states that a signal’s information is
preserved if it is uniformly sampled at a rate at least two times faster than its Fourier bandwidth.
Unfortunately, in many important and emerging applications, the resulting Nyquist rate can be
so high that we end up with too many samples and must compress in order to store or transmit
them. In other applications the cost of signal acquisition is prohibitive, either because of a high
cost per sample, or because state-of-the-art samplers cannot achieve the high sampling rates
required by Shannon/Nyquist. Examples include radar imaging and exotic imaging modalities
outside visible wavelengths.
Transform compression systems reduce the effective dimensionality of an N-dimensional
signal x by re-representing it in terms of a sparse set of coefficients α in a basis expansion
x = Ψα, with Ψ an N × N basis matrix. By sparse we mean that only K ≪ N of the
coefficients α are nonzero and need to be stored or transmitted. By compressible we mean that
the coefficients α, when sorted, decay rapidly enough to zero that α can be well-approximated as
K-sparse. The sparsity and compressibility properties are pervasive in many signals of interest.
For example, smooth signals and images are compressible in the Fourier basis, while piecewise
smooth signals and images are compressible in a wavelet basis [1]; the JPEG and JPEG2000
standards are examples of practical transform compression systems based on these bases.
Compressive sensing (CS) provides an alternative to Shannon/Nyquist sampling when the
signal under acquisition is known to be sparse or compressible [2–4]. In CS, we measure
not periodic signal samples but rather inner products with M ≪ N measurement vectors. In
matrix notation, the measurements y = Φx = ΦΨα, where the rows of the M × N matrix
Φ contain the measurement vectors. While the matrix ΦΨ is rank deficient, and hence loses
information in general, it can be shown to preserve the information in sparse and compressible
signals if it satisfies the so-called restricted isometry property (RIP) [3]. Intriguingly, a large
class of random matrices have the RIP with high probability. To recover the signal from the
compressive measurements y, we search for the sparsest coefficient vector α that agrees with
2
the measurements. To date, research in CS has focused primarily on reducing both the number
of measurements M (as a function of N and K) and on increasing the robustness and reducing
the computational complexity of the recovery algorithm. Today’s state-of-the-art CS systems
can robustly recover K-sparse and compressible signals from just M = O (K log(N/K)) noisy
measurements using polynomial-time optimization solvers or greedy algorithms.
While this represents significant progress from Nyquist-rate sampling, our contention in this
paper is that it is possible to do even better by more fully leveraging concepts from state-of-the-
art signal compression and processing algorithms. In many such algorithms, the key ingredient is
a more realistic signal model that goes beyond simple sparsity by codifying the inter-dependency
structure among the signal coefficients α.1 For instance, JPEG2000 and other modern wavelet
image coders exploit not only the fact that most of the wavelet coefficients of a natural image
are small but also the fact that the values and locations of the large coefficients have a particular
structure. Coding the coefficients according to a model for this structure enables these algorithms
to compress images close to the maximum amount possible – significantly better than a naıve
coder that just processes each large coefficient independently.
In this paper, we introduce a model-based CS theory that parallels the conventional theory and
provides concrete guidelines on how to create model-based recovery algorithms with provable
performance guarantees. By reducing the degrees of freedom of a sparse/compressible signal
by permitting only certain configurations of the large and zero/small coefficients, signal models
provide two immediate benefits to CS. First, they enable us to reduce, in some cases significantly,
the number of measurements M required to stably recover a signal. Second, during signal
recovery, they enable us to better differentiate true signal information from recovery artifacts,
which leads to a more robust recovery.
To precisely quantify the benefits of model-based CS, we introduce and study several new
theoretical concepts that could be of more general interest. We begin with signal models for K-
sparse signals and make precise how the structure encoded in a signal model reduces the number
1Obviously, sparsity and compressibility correspond to simple signal models where each coefficient is treated independently;
for example in a sparse model, the fact that the coefficient αi is large has no bearing on the size of any αj , j 6= i. We will
reserve the use of the term “model” for situations where we are enforcing dependencies between the values and the locations
of the coefficients αi.
3
of potential sparse signal supports in α. Then using the model-based restricted isometry property
(RIP) from [5, 6], we prove that such model-sparse signals can be robustly recovered from noisy
compressive measurements. Moreover, we quantify the required number of measurements M
and show that for some models M is independent of N . These results unify and generalize
the limited related work to date on signal models for strictly sparse signals [5–9]. We then
introduce the notion of a model-compressible signal, whose coefficients α are no longer strictly
sparse but have a structured power-law decay. To establish that model-compressible signals can
be robustly recovered from compressive measurements, we generalize the CS RIP to a new
restricted amplification property (RAmP). For some compressible signal models, the required
number of measurements M is independent of N .
To take practical advantage of this new theory, we demonstrate how to integrate signal models
into two state-of-the-art CS recovery algorithms, CoSaMP [10] and iterative hard thresholding
(IHT) [11]. The key modification is surprisingly simple: we merely replace the nonlinear ap-
proximation step in these greedy algorithms with a model-based approximation. Thanks to our
new theory, both new model-based recovery algorithms have provable robustness guarantees for
both model-sparse and model-compressible signals.
To validate our theory and algorithms and demonstrate its general applicability and utility, we
present two specific instances of model-based CS and conduct a range of simulation experiments.
The first model accounts for the fact that the large wavelet coefficients of piecewise smooth
signals and images tend to live on a rooted, connected tree structure [12]. Using the fact that the
number of such trees is much smaller than(
NK
), the number of K-sparse signal supports in N
dimensions, we prove that a tree-based CoSaMP algorithm needs only M = O (K) measurements
to robustly recover tree-sparse and tree-compressible signals. Figure 1 indicates the potential
performance gains on a tree-compressible, piecewise smooth signal.
The second model accounts for the fact that the large coefficients of many sparse signals cluster
together [7, 8]. Such a so-called block sparse model is equivalent to a joint sparsity model for
an ensemble of J , length-N signals [9], where the supports of the signals’ large coefficients are
shared across the ensemble. Using the fact that the number of clustered supports is much smaller
than(
JNK
), we prove that a block-based CoSaMP algorithm needs only M = O
While many natural and manmade signals and images can be described to first-order as sparse
or compressible, the support of their large coefficients often has an underlying inter-dependency
structure. This phenomenon has received only limited attention by the CS community to date [5–
8, 13–15]. In this section, we introduce a model-based theory of CS that captures such structure.
A model reduces the degrees of freedom of a sparse/compressible signal by permitting only
certain configurations of supports for the large coefficient. As we will show, this allows us to
reduce, in some cases significantly, the number of compressive measurements M required to
stably recover a signal.
A. Model-sparse signals
Recall from Section II-A that a K-sparse signal vector x lives in ΣK ⊂ RN , which is a union
of(
NK
)subspaces of dimension K. Other than its K-sparsity, there are no further constraints
on the support or values of its coefficients. A signal model endows the K-sparse signal x with
additional structure that allows certain K-dimensional subspaces in ΣK and disallows others [5,
6].
To state a formal definition of a signal model, let x|Ω represent the entries of x corresponding
to the set of indices Ω ⊆ 1, . . . , N, and let ΩC denote the complement of the set Ω.
Definition 2: A signal model MK is defined as the union of mK canonical K-dimensional
subspaces
MK =
mK⋃
m=1
Xm, such that Xm := x : x|Ωm ∈ RK , x|ΩC
m= 0,
where each subspace Xm contains all signals x with supp(x) ∈ Ωm. Thus, the model MK is
defined by the set of possible supports Ω1, . . . ,ΩmK.
Signals from MK are called K-model sparse. Clearly, MK ⊆ ΣK and contains mK ≤(
NK
)
subspaces.
10
In Sections V and VI below we consider two concrete models for sparse signals. The first
model accounts for the fact that the large wavelet coefficients of piecewise smooth signals and
images tend to live on a rooted, connected tree structure [12]. The second model accounts for
the fact that the large coefficients of sparse signals often cluster together [7–9].
B. Model-based RIP
If we know that the signal x being acquired is K-model sparse, then we can relax the RIP
constraint on the CS measurement matrix Φ and still achieve stable recovery from the compressive
measurements y = Φx [5, 6].
Definition 3: [5, 6] An M×N matrix Φ has theMK-restricted isometry property (MK-RIP)
with constant δMKif, for all x ∈MK , we have
(1− δMK)‖x‖22 ≤ ‖Φx‖22 ≤ (1 + δMK
)‖x‖22. (10)
To obtain a performance guarantee for model-based recovery of K-model sparse signals in
additive measurement noise, we must define an enlarged union of subspaces that includes sums
of elements in the model.
Definition 4: The B-Minkowski sum for the set MK , with B > 1 an integer, is defined as
MBK =
x =
B∑
r=1
x(r), with x(r) ∈MK
.
Define MB(x,K) as the algorithm that obtains the best approximation of x in the enlarged
union of subspaces MBK :
MB(x,K) = arg minx∈MB
K
‖x− x‖2.
We write M(x,K) := M1(x,K) when B = 1. Note that for many models, we will have
MBK ⊂MBK , and so the algorithm M(x,BK) will provide a strictly better approximation than
MB(x,K).
Our performance guarantee for model-sparse signal recovery will require that the measurement
matrix Φ be a near-isometry for all subspaces in MBK for some B > 1. This requirement is a
direct generalization of the 2K-RIP, 3K-RIP, and higher-order RIPs from the conventional CS
theory.
11
Blumensath and Davies [5] have quantified the number of measurements M necessary for a
random CS matrix to have the MK-RIP with a given probability.
Theorem 1: [5] Let MK be the union of mK subspaces of K-dimensions in RN . Then, for
any t > 0 and any
M ≥ 2
cδ2MK
(ln(2mK) +K ln
12
δMK
+ t
),
an M×N i.i.d. subgaussian random matrix has theMK-RIP with constant δMKwith probability
at least 1− e−t.
This bound can be used to recover the conventional CS result by substituting mK =(
NK
)≈
(Ne/K)K . The MK-RIP property is sufficient for robust recovery of model-sparse signals, as
we show below in Section IV-B.
C. Model-compressible signals
Just as compressible signals are “nearly K-sparse” and thus live close to the union of
subspaces ΣK in RN , model-compressible signals are “nearly K-model sparse” and live close
to the restricted union of subspaces MK . In this section, we make this new concept rigorous.
Recall from (3) that we defined compressible signals in terms of the decay of their K-term
approximation error.
The ℓ2 error incurred by approximating x ∈ RN by the best model-based approximation in
MK is given by
σMK(x) := inf
x∈MK
‖x− x‖2 = ‖x−M(x,K)‖2.
The decay of this approximation error defines the model-compressibility of a signal.
Definition 5: The set of s-model-compressible signals is defined as
Ms =x ∈ R
N : σMK(x) ≤ SK−1/s, 1 ≤ K ≤ N, S <∞
.
Define |x|Ms as the smallest value of S for which this condition holds for x and s.
We say that x ∈ Ms is an s-model-compressible signal under the signal model MK . These
approximation classes have been characterized for certain signal models; see Section V for an
example.
12
D. Nested model approximations and residual subspaces
In conventional CS, the same requirement (RIP) is a sufficient condition for the stable
recovery of both sparse and compressible signals. In model-based recovery, however, the class
of compressible signals is much larger than that of sparse signals, since the set of subspaces
containing model-sparse signals does not span all K-dimensional subspaces. Therefore, we need
to introduce some additional tools to develop a sufficient condition for the stable recovery of
model-compressible signals.
We will pay particular attention to models MK that generate nested approximations, since
they are more amenable to analysis.
Definition 6: A model M = M1,M2, . . . has the nested approximation property (NAP) if
supp(M(x,K)) ⊂ supp(M(x,K ′)) for all K < K ′ and for all x ∈ RN .
In words, a model generates nested approximations if the support of the best K ′-term model-
based approximation contains the support of the best K-term model-based approximation for all
K < K ′. An important example of a NAP model is the standard compressible signal model of
(3).
When a model obeys the NAP, the support of the difference between the best jK-term model-
based approximation and the best (j+ 1)K-term model-based approximation of a signal can be
shown to lie in a small union of subspaces, thanks to the structure enforced by the model. This
structure is captured by the set of subspaces that are included in each subsequent approximation,
as defined below.
Definition 7: The jth set of residual subspaces of size K is defined as
Rj,K(M) =u ∈ R
N such that u = M(x, jK)−M(x, (j − 1)K) for some x ∈ RN,
for j = 1, . . . , ⌈N/K⌉.Under the NAP, each signal x in a model can be partitioned into its best K-term approximation
xT1 , the additional components present in the best 2K-term approximation xT2 , and so on, with
x =∑⌈N/K⌉
j=1 xTjand xTj
∈ Rj,K(M) for each j. Each signal partition xTjis a K-sparse signal,
and thus Rj,K(M) is a union of subspaces of dimension K. We will denote by Rj the number
of subspaces that compose Rj,K(M) and omit the dependence on M in the sequel for brevity.
13
Intuitively, the norms of the partitions ‖xTj‖2 decay as j increase for signals that are
compressible under the model. As the next subsection shows, this observation is instrumental in
relaxing the isometry restrictions on the measurement matrix Φ and bounding the recovery error
for s-model-compressible signals when the model obeys the NAP.
E. The restricted amplification property (RAmP)
For exactly K-model-sparse signals, we discussed in Section III-B that the number of
compressive measurements M required for a random matrix to have theMK-RIP is determined
by the number of canonical subspaces mK via (11). Unfortunately, such model-sparse concepts
and results do not immediately extend to model-compressible signals. Thus, we develop a
generalization of the MK-RIP that we will use to quantify the stability of recovery for model-
compressible signals.
One way to analyze the robustness of compressible signal recovery in conventional CS is
to consider the tail of the signal outside its K-term approximation as contributing additional
“noise” to the measurements of size ‖Φ(x− xK)‖2 [10, 11, 23]. Consequently, the conventional
K-sparse recovery performance result can be applied with the augmented noise n+ Φ(x−xK).
This technique can also be used to quantify the robustness of model-compressible signal
recovery. The key quantity we must control is the amplification of the model-based approximation
residual through Φ. The following property is a new generalization of the RIP and model-based
RIP.
Definition 8: A matrix Φ has the (ǫK , r)-restricted amplification property (RAmP) for the
residual subspaces Rj,K of model M if
‖Φu‖22 ≤ (1 + ǫK)j2r‖u‖22 (11)
for any u ∈ Rj,K for each 1 ≤ j ≤ ⌈N/K⌉.The regularity parameter r > 0 caps the growth rate of the amplification of u ∈ Rj,K as a
function of j. Its value can be chosen so that the growth in amplification with j balances the
decay of the norm in each residual subspace Rj,K with j.
We can quantify the number of compressive measurements M required for a random
measurement matrix Φ to have the RAmP with high probability; we prove the following in
14
Appendix I.
Theorem 2: Let Φ be an M×N matrix with i.i.d. subgaussian entries and let the set of residual
subspaces Rj,K of model M contain Rj subspaces of dimension K for each 1 ≤ j ≤ ⌈N/K⌉.If
M ≥ max1≤j≤⌈N/K⌉
1(jr√
1 + ǫK − 1)2(
2K + 4 lnRjN
K+ 2t
), (12)
then the matrix Φ has the (ǫK , r)-RAmP with probability 1− e−t.
The order of the bound of Theorem 2 is lower than O (K log(N/K)) as long as the number of
subspaces Rj grows slower than NK . Armed with the RaMP, we now prove the following result,
which will provide robustness for the recovery of model-compressible signals; see Appendix II
for the proof.
Theorem 3: Let x ∈Ms be an s-model compressible signal under a model M that obeys the
NAP. If Φ has the (ǫK , r)-RAmP and r = s− 1, then we have
‖Φ(x−M(x,K))‖2 ≤√
1 + ǫKK−s ln
⌈N
K
⌉|x|Ms .
IV. MODEL-BASED SIGNAL RECOVERY ALGORITHMS
To take practical advantage of our new theory for model-based CS, we demonstrate how to
integrate signal models into two state-of-the-art CS recovery algorithms, CoSaMP [10] (in this
section) and iterative hard thresholding (IHT) [11] (in Appendix III). The key modification is
simple: we merely replace the best K-term approximation step in these greedy algorithms with a
best K-term model-based approximation. Since at each iteration we need only search over the mK
subspaces ofMK rather than(
NK
)subspaces of ΣK , fewer measurements will be required for the
same degree of robust signal recovery. Or, alternatively, using the same number of measurements,
more accurate recovery can be achieved. After presenting the modified CoSaMP algorithm, we
prove robustness guarantees for both model-sparse and model-compressible signals.
A. Model-based CoSaMP
We choose to modify the CoSaMP algorithm [10] for two reasons. First, it has robust recovery
guarantees that are on par with the best convex optimization-based approaches. Second, it
15
Algorithm 1 Model-based CoSaMP
Inputs: CS matrix Φ, measurements y, model MK
Output: K-sparse approximation x to true signal x
x0 = 0, r = y, i = 0 initializewhile halting criterion false do
i← i+ 1
e← ΦT r form signal residual estimateΩ← supp(M2(e,K)) prune signal residual estimate according to signal modelT ← Ω ∪ supp(xi−1) merge supportsb|T ← Φ†
T y, b|T C ← 0 form signal estimatexi ←M(b,K) prune signal estimate according to signal modelr ← y − Φxi update measurement residual
end while
return x← xi
has a simple iterative, greedy structure based on a best BK-term approximation (with B a
small integer) that is easily modified to incorporate a best BK-term model-based approximation
MB(K, x). Pseudocode for the modified algorithm is given in Algorithm 1.
We now study the performance of model-based CoSaMP signal recovery on model-sparse
signals and model-compressible signals.
B. Performance of model-sparse signal recovery
A robustness guarantee for noisy measurements of model-sparse signals can be obtained using
the model-based RIP (10). The following theorem is proven in Appendix IV.
Theorem 4: Let x ∈ MK and let y = Φx + n be a set of noisy CS measurements. If Φ has
an M4K-RIP constant of δM4
K≤ 0.1, then the signal estimate xi obtained from iteration i of the
model-based CoSaMP algorithm satisfies
‖x− xi‖2 ≤ 2−i‖x‖2 + 15‖n‖2. (13)
16
C. Performance of model-compressible signal recovery
Using the new tools introduced in Section III, we can provide a robustness guarantee for
noisy measurements of model-compressible signals, using the RAmP as a condition on the
measurement matrix Φ.
Theorem 5: Let x ∈Ms be an s-model-compressible signal from a model M that obeys the
NAP, and let y = Φx + n be a set of noisy CS measurements. If Φ has the M4K-RIP with
δM4K≤ 0.1 and the (ǫK , r)-RAmP with ǫK ≤ 0.1 and r = s − 1, then the signal estimate xi
obtained from iteration i of the model-based CoSaMP algorithm satisfies
‖x− xi‖2 ≤ 2−i‖x‖2 + 35(‖n‖2 + |x|MsK
−s(1 + ln⌈N/K⌉)). (14)
To prove the theorem, we first bound the recovery error for an s-model-compressible signal
x ∈ Ms when the matrix Φ has the (ǫK , r)-RAmP with r ≤ s − 1. Then, using Theorems 3
and 4, we can easily prove the result by following the analogous proof in [10].
D. Robustness to model mismatch
We now analyze the robustness of model-based CS recovery to model mismatch, which occurs
when the signal being recovered from compressive measurements does not conform exactly to
the model used in the recovery algorithm.
We begin with optimistic results for signals that are “close” to matching the recovery model.
First consider a signal x that is not K-model sparse as the recovery algorithm assumes but rather
(K + κ)-model sparse for some small integer κ. This signal can be decomposed into xK , the
signal’s K-term model-based approximation, and x − xK , the error of this approximation. For
κ ≤ K, we have that x−xK ∈ R2,K . If the matrix Φ has the (ǫK , r)-RAmP, then it follows than
‖Φ(x− xK)‖2 ≤ 2r√
1 + ǫK‖x− xK‖2. (15)
Using equations (13) and (15), we obtain the following guarantee for the ith iteration of model-
based CoSaMP:
‖x− xi‖2 ≤ 2−i‖x‖2 + 16 · 2r√
1 + ǫK‖x− xK‖2 + 15‖n‖2.
17
By noting that ‖x− xK‖2 is small, we obtain a guarantee that is close to (13).
Second, consider a signal x that is not s-model compressible as the recovery algorithm assumes
but rather (s−ǫ)-model compressible. The following bound can be obtained under the conditions
of Theorem 5 by modifying the argument in Appendix II:
‖x− xi‖2 ≤ 2−i‖x‖2 + 35
(‖n‖2 + |x|MsK
−s
(1 +⌈N/K⌉ǫ − 1
ǫ
)).
As ǫ becomes smaller, the factor⌈N/K⌉ǫ−1
ǫapproaches log⌈N/K⌉, matching (14). In summary,
as long as the deviations from the model-sparse and model-compressible models are small, our
model-based recovery guarantees still apply within a small bounded constant factor.
We end with a more pessimistic, worst-case result for signals that are arbitrarily far away
from model-sparse or model-compressible. Consider such an arbitrary x ∈ RN and compute its
nested model-based approximations xjK = M(x, jK), j = 1, . . . , ⌈N/K⌉. If x is not model-
compressible, then the model-based approximation error σjK(x) is not guaranteed to decay as j
decreases. Additionally, the number of residual subspaces Rj,K could be as large as(
NK
); that is,
the jth difference between subsequent model-based approximations xTj= xjK − x(j−1)K might
lie in any arbitrary K-dimensional subspace. This worst case is equivalent to setting r = 0 and
Rj =(
NK
)in Theorem 2. It is easy to see that this condition on the number of measurements
M is nothing but the standard RIP for CS. Hence, if inflate the number of measurements to
M = O (K log(N/K)) (the usual number for conventional CS), the performance of model-based
CoSaMP recovery on an arbitrary signal x follows the K-term model-based approximation of x
within a bounded constant factor.
E. Computational complexity of model-based recovery
The computational complexity of a model-based signal recovery algorithm differs from
that of a standard algorithm by two factors. The first factor is the reduction in the number
of measurements M necessary for recovery: since most current recovery algorithms have a
computational complexity that is linear in the number of measurements, any reduction in M
reduces the total complexity. The second factor is the cost of the model-based approximation.
The K-term approximation used in most current recovery algorithms can be implemented with a
18
simple sorting operation (O (N logN) complexity, in general). Ideally, the signal model should
support a similarly efficient approximation algorithm.
To validate our theory and algorithms and demonstrate their general applicability and utility,
we now present two specific instances of model-based CS and conduct a range of simulation
experiments.
V. EXAMPLE: WAVELET TREE MODEL
Wavelet decompositions have found wide application in the analysis, processing, and com-
pression of smooth and piecewise smooth signals because these signals are K-sparse and
compressible, respectively [1]. Moreover, the wavelet coefficients can be naturally organized
into a tree structure, and for many kinds of natural and manmade signals the largest coefficients
cluster along the branches of this tree. This motivates a connected tree model for the wavelet
coefficients [24–26].
While CS recovery for wavelet-sparse signals has been considered previously [13–15],
the resulting algorithms integrated the tree constraint in an ad-hoc fashion. Furthermore, the
algorithms provide no recovery guarantees or bounds on the necessary number of compressive
measurements.
A. Tree-sparse signals
We first describe tree sparsity in the context of sparse wavelet decompositions. We focus
on one-dimensional signals and binary wavelet trees, but all of our results extend directly to
d-dimensional signals and 2d-ary wavelet trees.
Consider a signal x of length N = 2I , for an integer value of I . The wavelet representation
of x is given by
x = v0ν +
I−1∑
i=0
2i−1∑
j=0
wi,jψi,j ,
where ν is the scaling function and ψi,j is the wavelet function at scale i and offset j. The
wavelet transform consists of the scaling coefficient v0 and wavelet coefficients wi,j at scale i,
0 ≤ i ≤ I − 1, and position j, 0 ≤ j ≤ 2i − 1. In terms of our earlier matrix notation, x has
the representation x = Ψα, where Ψ is a matrix containing the scaling and wavelet functions as
19
...Fig. 2. Binary wavelet tree for a one-dimensional signal. The squares denote the large wavelet coefficients that arise
from the discontinuities in the piecewise smooth signal drawn below; the support of the large coefficients forms a
rooted, connected tree.
columns, and α = [v0 w0,0 w1,0 w1,1 w2,0 . . .]T is the vector of scaling and wavelet coefficients.
We are, of course, interested in sparse and compressible α.
The nested supports of the wavelets at different scales create a parent/child relationship
between wavelet coefficients at different scales. We say that wi−1,⌊j/2⌋ is the parent of wi,j
and that wi+1,2j and wi+1,2j+1 are the children of wi,j. These relationships can be expressed
graphically by the wavelet coefficient tree in Figure 2.
Wavelet functions act as local discontinuity detectors, and using the nested support property
of wavelets at different scales, it is straightforward to see that a signal discontinuity will give
rise to a chain of large wavelet coefficients along a branch of the wavelet tree from a leaf to
the root. Moreover, smooth signal regions will give rise to regions of small wavelet coefficients.
This “connected tree” property has been well-exploited in a number of wavelet-based processing
[12, 27, 28] and compression [29, 30] algorithms. In this section, we will specialize the theory
developed in Sections III and IV to a connected tree model T .
A set of wavelet coefficients Ω forms a connected subtree if, whenever a coefficient wi,j ∈ Ω,
then its parent wi−1,⌊j/2⌋ ∈ Ω as well. Each such set Ω defines a subspace of signals whose
support is contained in Ω; that is, all wavelet coefficients outside Ω are zero. In this way, we
define the model TK as the union of all K-dimensional subspaces corresponding to supports Ω
that form connected subtrees.
20
Definition 9: Define the set of K-tree sparse signals as
TK =
x = v0ν +
I−1∑
i=0
2i∑
j=1
wi,jψi,j : w|ΩC = 0, |Ω| = K,Ω forms a connected subtree
.
To quantify the number of subspaces in TK , it suffices to count the number of distinct connected
subtrees of size K in a binary tree of size N . We prove the following result in Appendix V.
Proposition 1: The number of subspaces in TK obeys TK ≤ 4K+4
Ke2 for K ≥ log2N and
TK ≤ (2e)K
K+1for K < log2N .
B. Tree-based approximation
To implement tree-based signal recovery, we seek an efficient algorithm T(x,K) to solve the
optimal approximation
xTK = arg minx∈TK
‖x− x‖2. (16)
Fortuitously, an efficient solver exists, called the condensing sort and select algorithm (CSSA)
[24–26]. Recall that subtree approximation coincides with standard K-term approximation (and
hence can be solved by simply sorting the wavelet coefficients) when the wavelet coefficients
are monotonically nonincreasing along the tree branches out from the root. The CSSA solves
(16) in the case of general wavelet coefficient values by condensing the nonmonotonic segments
of the tree branches using an iterative sort-and-average routine. The condensed nodes are called
“supernodes”. Condensing a large coefficient far down the tree accounts for the potentially large
cost (in terms of the total budget of tree nodes K) of growing the tree to that point.
The CSSA can also be interpreted as a greedy search among the nodes. For each node in the
tree, the algorithm calculates the average wavelet coefficient magnitude for each subtree rooted
at that node, and records the largest average among all the subtrees as the energy for that node.
The CSSA then searches for the unselected node with the largest energy and adds the subtree
corresponding to the node’s energy to the estimated support as a supernode [26].
Since the first step of the CSSA involves sorting all of the wavelet coefficients, overall it
requires O (N logN) computations. However, once the CSSA grows the optimal tree of size K,
it is trivial to determine the optimal trees of size < K and computationally efficient to grow the
optimal trees of size > K [24].
21
The constrained optimization (16) can be rewritten as an unconstrained problem by introducing
the Lagrange multiplier λ [31]:
minx∈T‖x− x‖22 + λ(‖α‖0 −K),
where T = ∪Nn=1Tn and α are the wavelet coefficients of x. Except for the inconsequential
λK term, this optimization coincides with Donoho’s complexity penalized sum of squares [31],
which can be solved in only O (N) computations using coarse-to-fine dynamic programming on
the tree. Its primary shortcoming is the nonobvious relationship between the tuning parameter
λ and and the resulting size K of the optimal connected subtree.
C. Tree-compressible signals
Specializing Definition 2 from Section III-C to T , we make the following definition.
Definition 10: Define the set of s-tree compressible signals as
Ts =x ∈ R
N : ‖x− T(x,K)‖2 ≤ SK−s, 1 ≤ K ≤ N, S <∞.
Furthermore, define |x|Ts as the smallest value of S for which this condition holds for x and s.
Tree approximation classes contain signals whose wavelet coefficients have a loose (and
possibly interrupted) decay from coarse to fine scales. These classes have been well-characterized
for wavelet-sparse signals [25, 26, 30] and are intrinsically linked with the Besov spaces
Bsq(Lp([0, 1])). Besov spaces contain functions of one or more continuous variables that have
(roughly speaking) s derivatives in Lp([0, 1]); the parameter q provides finer distinctions of
smoothness. When a Besov space signal xa with s > 1/p − 1/2 is sampled uniformly and
converted to a length-N vector x, its wavelet coefficients belong to the tree approximation space
Ts, with
|xN |Ts ≍ ‖xa‖Lp([0,1]) + ‖xa‖Bsq (Lp([0,1])),
where “≍” denotes an equivalent norm. The same result holds if s = 1/p− 1/2 and q ≤ p.
D. Stable tree-based recovery from compressive measurements
For tree-sparse signals, by applying Theorem 1 and Proposition 1, we find that a subgaussian
random matrix has the TK-RIP property with constant δTKand probability 1−e−t if the number
22
of measurements obeys
M ≥
2cδ2
TK
(K ln 48
δTK
+ ln 512Ke2 + t
)if K < log2N,
2cδ2
TK
(K ln 24e
δTK+ ln 2
K+1+ t)
if K ≥ log2N,
Thus, the number of measurements necessary for stable recovery of tree-sparse signals is linear
in K, without the dependence on N present in conventional non-model-based CS recovery.
For tree-compressible signals, we must quantify the number of subspaces Rj in each residual
set Rj,K for the approximation class. We can then apply the theory of Section IV-C with
Proposition 1 to calculate smallest allowable M via Theorem 5.
Proposition 2: The number of K-dimensional subspaces that comprise Rj,K obeys
Rj ≤
(2e)K(2j+1)
(Kj+K+1)(Kj+1)if 1 ≤ j <
⌊log2 N
K
⌋,
2(3j+2)K+8ejK
(Kj+1)K(j+1)e2 if j =⌊
log2 NK
⌋,
4(2j+1)K+8
K2j(j+1)e4 if j >⌊
log2 NK
⌋.
(17)
Using Proposition 2 and Theorem 5, we obtain the following condition for the matrix Φ to
have the RAmP, which is proved in Appendix VI.
Proposition 3: Let Φ be an M ×N matrix with i.i.d. subgaussian entries.. If
M ≥
2
(√
1+ǫK−1)2
(10K + 2 ln N
K(K+1)(2K+1)+ t)
if K ≤ log2N,
2
(√
1+ǫK−1)2
(10K + 2 ln 601N
K3 + t)
if K > log2N,
then the matrix Φ has the (ǫK , s)-RAmP for model T and all s > 0.5 with probability 1− e−t.
Both cases give a simplified bound on the number of measurements required as M = O (K),
which is a substantial improvement over the M = O (K log(N/K)) required by conventional
CS recovery methods. Thus, when Φ satisfies Proposition 3, we have the guarantee (14) for
sampled Besov space signals from Bsq(Lp([0, 1])).
E. Experiments
We now present the results of a number of numerical experiments that illustrate the
effectiveness of a tree-based recovery algorithm. Our consistent observation is that explicit
incorporation of the model in the recovery process significantly improves the quality of recovery
23
for a given number of measurements. In addition, model-based recovery remains stable when the
inputs are no longer tree-sparse, but rather are tree-compressible and/or corrupted with differing
levels of noise. We employ the model-based CoSaMP recovery of Algorithm 1 with a CSSA-
based approximation step in all experiments.
We first study one-dimensional signals that match the connected wavelet-tree model described
above. Among such signals is the class of piecewise smooth functions, which are commonly
encountered in analysis and practice.
Figure 1 illustrates the results of recovering the tree-compressible HeaviSine signal of length
N = 1024 from M = 80 noise-free random Gaussian measurements using CoSaMP, ℓ1-norm
minimization using the l1 eq solver from the ℓ1-Magic toolbox,3 and our tree-based recovery
algorithm. It is clear that the number of measurements (M = 80) is far fewer than the minimum
number required by CoSaMP and ℓ1-norm minimization to accurately recover the signal. In
contrast, tree-based recovery using K = 26 is accurate and uses fewer iterations to converge
than conventional CoSaMP. Moreover, the normalized magnitude of the squared error for tree-
based recovery is equal to 0.037, which is remarkably close to the error between the noise-free
signal and its best K-term tree-approximation (0.036).
Figure 3 illustrates the results of a Monte Carlo simulation study on the impact of the number
of measurements M on the performance of model-based and conventional recovery for a class
of tree-sparse piecewise-polynomial signals. Each data point was obtained by measuring the
normalized recovery error of 500 sample trials. Each sample trial was conducted by generating
a new piecewise-polynomial signal with five polynomial pieces of cubic degree and randomly
placed discontinuities, computing its best K-term tree-approximation using the CSSA, and then
measuring the resulting signal using a matrix with i.i.d. Gaussian entries. Model-based recovery
attains near-perfect recovery at M = 3K measurements, while CoSaMP only matches this
performance at M = 5K. We defer a full Monte Carlo comparison of our method with the
much more computationally demanding ℓ1-norm minimization to future work. In practice, we
have noticed that CoSaMP and ℓ1-norm minimization offer similar recovery trends; consequently,
we can expect that model-based recovery will offer a similar degree of improvement over ℓ1-norm
3http://www.acm.caltech.edu/l1magic.
24
2 2.5 3 3.5 4 4.5 50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
M/KA
vera
ge n
orm
aliz
ed e
rror
mag
nitu
de Model−based recoveryCoSaMP
Fig. 3. Performance of CoSaMP vs. wavelet tree-based recovery on a class of piecewise-cubic signals as a function
of M/K .
minimization.
Further, we demonstrate that model-based recovery performs stably in the presence of
measurement noise. We generated sample piecewise-polynomial signals as above, computed
their best K-term tree-approximations, computed M measurements of each approximation, and
finally added Gaussian noise of expected norm ‖n‖2 to each measurement. Then, we recovered
the signal using CoSaMP and model-based recovery and measured the recovery error in each case.
For comparison purposes, we also tested the recovery performance of a ℓ1-norm minimization
algorithm that accounts for the presence of noise, which has been implemented as the l1 qc
solver in the ℓ1-Magic toolbox. First, we determined the lowest value of M for which the
respective algorithms provided near-perfect recovery in the absence of noise in the measurements.
This corresponds to M = 3.5K for model-based recovery, M = 5K for CoSaMP, and M = 4.5K
for ℓ1 minimization. Next, we generated 200 sample tree-modeled signals, computed M noisy
measurements, recovered the signal using the given algorithm and recorded the recovery error.
Figure 4 illustrates the growth in maximum normalized recovery error (over the 200 sample
trials) as a function of the expected measurement signal-to-noise ratio for the tree algorithms. We
observe similar stability curves for all three algorithms, while noting that model-based recovery
offers this kind of stability using significantly fewer measurements.
Finally, we turn to two-dimensional images and a wavelet quadtree model. The connected
25
0 0.1 0.2 0.3 0.4 0.50
0.2
0.4
0.6
0.8
1CoSaMP (M = 5K)
))
Max
imum
nor
mal
ized
rec
over
y er
ror
Fig. 4. Robustness to measurement noise for standard and wavelet tree-based CS recovery algorithms. We plot the
maximum normalized recovery error over 200 sample trials as a function of the expected signal-to-noise ratio. The
linear growth demonstrates that model-based recovery possesses the same robustness to noise as CoSaMP and ℓ1-norm
minimization.
wavelet-tree model has proven useful for compressing natural images [25]; thus, our algorithm
provides a simple and provably efficient method for recovering a wide variety of natural images
from compressive measurements. An example of recovery performance is given in Figure 5. The
test image (Peppers) is of size N = 128 × 128 = 16384 pixels, and we computed M = 5000
random Gaussian measurements. Model-based recovery again offers higher performance than
standard signal recovery algorithms like CoSaMP, both in terms of recovery mean-squared error
and visual quality.
VI. EXAMPLE: BLOCK-SPARSE SIGNALS AND SIGNAL ENSEMBLES
In a block-sparse signal, the locations of the significant coefficients cluster in blocks under
a specific sorting order. Block-sparse signals have been previously studied in CS applications,
including DNA microarrays and magnetoencephalography [7, 8]. An equivalent problem arises
in CS for signal ensembles, such as sensor networks and MIMO communication [8, 9, 32]. In this
case, several signals share a common coefficient support set. For example, when a frequency-
sparse acoustic signal is recorded by an array of microphones, then all of the recorded signals
contain the same Fourier frequencies but with different amplitudes and delays. Such a signal
26
(a) Peppers (b) CoSaMP (c) model-based recovery
(RMSE = 22.8) (RMSE = 11.1)
Fig. 5. Example performance of standard and model-based recovery on images. (a) N = 128× 128 = 16384-pixel
Peppers test image. Image recovery from M = 5000 compressive measurements using (b) conventional CoSaMP and
(c) our wavelet tree-based algorithm.
ensemble can be re-shaped as a single vector by concatenation, and then the coefficients can be
rearranged so that the concatenated vector exhibits block sparsity.
It has been shown that the block-sparse structure enables signal recovery from a reduced
number of CS measurements, both for the single signal case [7, 8] and the signal ensemble
case [9], through the use of specially tailored recovery algorithm [7, 8, 33]. However, the
robustness guarantees for such algorithms either are restricted to exactly sparse signals and
noiseless measurements, do not have explicit bounds on the number of necessary measurements,
or are asymptotic in nature.
In this section, we formulate the block sparsity signal model as a union of subspaces and
pose an approximation algorithm on this union of subspaces. The approximation algorithm is
used to implement block-based signal recovery. We also define the corresponding class of block-
compressible signals and quantify the number of measurements necessary for robust recovery.
A. Block-sparse signals
Consider a class S of signal vectors x ∈ RJN , with J and N integers. This signal can be
reshapped into a J × N matrix X , and we use both notations interchangeably in the sequel.
27
We will restrict entire columns of X to be part of the support of the signal as a group. That
is, signals X in a block-sparse model have entire columns as zeros or nonzeros. The measure
of sparsity for X is its number of nonzero columns. More formally, we make the following
definition.
Definition 11: [7, 8] Define the set of K-block sparse signals as
SK = X = [x1 . . . xN ] ∈ RJ×N such that xn = 0 for n /∈ Ω,Ω ⊆ 1, . . . , N, |Ω| = K.
It is important to note that a K-block sparse signal has sparsity KJ , which is dependent on
the size of the block J . We can extend this formulation to ensembles of J , length-N signals
with common support. Denote this signal ensemble by x1, . . . , xJ, with xj ∈ RN , 1 ≤ j ≤ J .
We formulate a matrix representation X of the ensemble that features the signal xj in its jth
row: X = [x1 . . . xN ]T . The matrix X features the same structure as the matrix X obtained
from a block-sparse signal; thus, the matrix X can be converted into a block-sparse vector x
that represents the signal ensemble.
B. Block-based approximation
To pose a the block-based approximation algorithm, we need to define the mixed norm of a
matrix.
Definition 12: The (p, q) mixed norm of the matrix X = [x1 x2 . . . xN ] is defined as
‖X‖(p,q) =
(N∑
n=1
‖xn‖qp
)1/q
.
When q = 0, ‖X‖(p,0) simply counts the number of nonzero columns in X .
We immediately find that ‖X‖(p,p) = ‖x‖p, with x the vectorization of X . Intuitively, we pose
the algorithm S(X,K) to obtain the best block-based approximation of the signal X as follows:
XSK = arg min
X∈RJ×N‖X − X‖(2,2) subject to ‖X‖(2,0) ≤ K. (18)
It is easy to show that to obtain the approximation, it suffices to perform column-wise hard
thresholding: let ρ be the K th largest ℓ2-norm among the columns of X . Then our approximation
algorithm is S(X,K) = XSK = [xSK,1 . . . xSK,N ], where
xSK,n =
xn ‖xn‖2 ≥ ρ,
0 ‖xn‖2 < ρ,
28
for each 1 ≤ j ≤ J and 1 ≤ n ≤ N . Alternatively, a recursive approximation algorithm can be
obtained by sorting the columns of X by their ℓ2 norms, and then selecting the largest columns.
The complexity of this sorting process is O (NJ +N logN).
C. Block-compressible signals
The approximation class under the block-compressible model corresponds to signals with
blocks whose ℓ2 norm has a power-law decay rate.
Definition 13: We define the set of s-block compressible signals as
Ss = X = [x1 . . . xN ] ∈ RJ×N s.t. ‖xI(i)‖2 ≤ Si−s−1/2, 1 ≤ i ≤ N, S <∞,
where I indexes the sorted column norms.
We say that X is an s-block compressible signal if X ∈ Ss. For such signals, we have ‖X −XK‖(2,2) = σSK
(x) ≤ S1K−s, and ‖X−XK‖(2,1) ≤ S2K
1/2−s. Note that the block-compressible
signal model does not impart a structure to the decay of the signal coefficients, so that the sets
Rj,K are equal for all values of j; due to this property, the (δSK, s)-RAmP is implied by the
SK-RIP. Taking this into account, we can derive the following result from [10], which is proven
similarly to Theorem 4.
Theorem 6: Let x be a signal from model S, and let y = Φx + n be a set of noisy CS
measurements. If Φ has the S4K-RIP with δS4
K≤ 0.1, then the estimate obtained from iteration i
of block-based CoSaMP, using the approximation algorithm (18), satisfies
‖x− xi‖2 ≤ 2−i‖x‖2 + 20
(‖X −XS
K‖(2,2) +1√K‖X −XS
K‖(2,1) + ‖n‖2).
Thus, the algorithm provides a recovered signal of similar quality to approximations of X
by a small number of nonzero columns. When the signal x is K-block sparse, we have that
||X − XSK‖(2,2) = ||X − XS
K‖(2,1) = 0, obtaining the same result as Theorem 4, save for a
constant factor.
D. Stable block-based recovery from compressive measurements
Since Theorem 6 poses the same requirement on the measurement matrix Φ for sparse and
compressible signals, the same number of measurements M is required to provide performance
29
guarantees for block-sparse and block-compressible signals. The class SK contains S =(
NK
)
subspaces of dimension JK. Thus, a subgaussian random matrix has the SK-RIP property with
constant δSKand probability 1− e−t if the number of measurements obeys
M ≥ 2
cδ2SK
(K
(ln
2N
K+ J ln
12
δSK
)+ t
). (19)
The first term in this bound matches the order of the bound for conventional CS, while the
second term introduces a linear dependence on the size of the block J . This shows that the
number of measurements required for robust recovery scales as M = O (KJ +K log(N/K)),
which is a substantial improvement over the M = O (JK log(N/K)) that would be required by
conventional CS recovery methods. When the size of the block J is larger than log(N/K), then
this term becomes O (KJ); that is, it is linear on the total sparsity of the block-sparse signal.
We note in passing that the bound on the number of measurements (19) assumes a dense
subgaussian measurement matrix, while the measurement matrices used in [9] have a block-
diagonal. structure. To obtain measurements from an M × JN dense matrix in a distributed
setting, it suffices to partition the matrix into J pieces of size M × N and calculate the CS
measurements at each sensor with a corresponding matrix; these individual measurements are
then summed to obtain the complete measurement vector. For large J , (19) implies that the
total number of measurements required for recovery of the signal ensemble is lower than the
bound for the case where each signal recovery is performed independently for each signal (M
= O (JK log(N/K))).
E. Experiments
We conducted several numerical experiments comparing model-based recovery to CoSaMP
in the context of block-sparse signals. We employ the model-based CoSaMP recovery of
Algorithm 1 with the block-based approximation algorithm (18) in all cases. For brevity, we
exclude a thorough comparison of our model-based algorithm with ℓ1-based optimization and
defer it to future work. In practice, we observed that our algorithm performs several times faster
than convex optimization-based procedures.
Figure 6 illustrates an N = 4096 signal that exhibits block sparsity, and its recovered version
using CoSaMP and model-based recovery. The block size J = 64 and there were K = 6 active
30
(a) original block-sparse signal (b) CoSaMP (c) model-based recovery
(RMSE = 0.723) (RMSE = 0.015)
Fig. 6. Example performance of model-based signal recovery for a block-sparse signal. (a) Example block-
compressible signal of length N = 4096 with K = 6 nonzero blocks of size J = 64. Recovered signal from
M = 960 measurements using (b) conventional CoSaMP recovery and (c) block-based recovery.
blocks in the signal. We observe the clear advantage of using the block-sparsity model in signal
recovery.
We now consider block-compressible signals. An example recovery is illustrated in Figure 7. In
this case, the ℓ2-norms of the blocks decay according to a power law, as described above. Again,
the number of measurements is far below the minimum number required to guarantee stable
recovery through conventional CS recovery. However, enforcing the model in the approximation
process results in a solution that is very close to the best 5-block approximation of the signal.
Figure 8 indicates the decay in recovery error as a function of the numbers of measurements
for CoSaMP and model-based recovery. We generated sample block-sparse signals as follows:
we randomly selected a set of K blocks, each of size J , and endow them with coefficients that
follow an i.i.d. Gaussian distribution. Each sample point in the curves is generated by performing
200 trials of the corresponding algorithm. As in the connected wavelet-tree case, we observe
clear gains using model-based recovery, particularly for low-measurement regimes; CoSaMP
matches model-based recovery only for M ≥ 5K.
VII. CONCLUSIONS
In this paper, we have aimed to demonstrate that there are significant performance gains to
be made by exploiting more realistic and richer signal models beyond the simplistic sparse
and compressible models that dominate the CS literature. Building on the unions of subspaces
31
(a) signal (b) best 5-block approximation
(RMSE = 0.116)
(c) CoSaMP (d) model-based recovery
(RMSE = 0.711) (RMSE = 0.195)
Fig. 7. Example performance of model-based signal recovery for block-compressible signals. (a) Example block-
compressible signal, length N = 1024. (b) Best block-based approximation with K = 5 blocks. Recovered signal
from M = 200 measurements using both (c) conventional CoSaMP recovery and (d) block-based recovery.
results of [5] and the proof machinery of [10], we have taken some of the first steps towards
what promises to be a general theory for model-based CS by introducing the notion of a model-
compressible signal and the associated restricted amplification property (RAmP) condition it
imposes on the measurement matrix Φ.
For the volumes of natural and manmade signals and images that are wavelet-sparse or
compressible, our tree-based CoSaMP and IHT algorithms offer performance that signifi-
cantly exceeds today’s state-of-the-art while requiring only M = O (K) rather than M =
O (K log(N/K)) random measurements. For block-sparse signals and signal ensembles, our
block-based CoSaMP and IHT algorithms offer not only excellent performance but also require
just M = O (JK) measurements, where JK is the signal sparsity. Furthermore, block-based
32
1 2 3 4 50
2
4
6
8
10
12
14
16
M/KA
vera
ge n
orm
aliz
ed e
rror
mag
nitu
de Model−based recoveryCoSaMP
Fig. 8. Performance of CoSaMP and block-based recovery on a class of block-sparse signals as a function of M/K .
Standard CS recovery does not match the performance of block-based recovery until M = 5K .
recovery can recovery signal ensembles using fewer measurements than the number required
when each signal is recovered independently.
There are many avenues for future work on model-based CS. We have only considered the
recovery of signals from models that can be geometrically described as a union of subspaces;
possible extensions include other, more complex geometries (for example, high-dimensional
polytopes, nonlinear manifolds.) We also expect that the core of our proposed algorithms — a
model-enforcing approximation step — can be integrated into other iterative algorithms, such
as relaxed ℓ1-norm minimization methods. Furthermore, our framework will benefit from the
formulation of new signal models that are endowed with efficient model-based approximation
algorithms.
ACKNOWLEDGEMENTS
We thank Petros Boufounos and Mark Davenport for helpful discussions.
APPENDIX I
PROOF OF THEOREM 2
To prove this theorem, we will study the distribution of the maximum singular value of a
submatrix ΦT of a matrix with i.i.d. Gaussian entries Φ corresponding to the columns indexed
33
by T . From this we obtain the probability that RAmP does not hold for a fixed support T . We
will then evaluate the same probability for all supports T of elements of Rj,K , where the desired
bound on the amplification is dependent on the value of j. This gives us the probability that the
RAmP does not hold for a given residual subspace set Rj,K . We fix the probability of failure
on each of these sets; we then obtain probability that the matrix Φ does not have the RAmP
using a union bound. We end by obtaining conditions on the number of rows M of Φ to obtain
a desired probability of failure.
We begin from the following concentration of measure for the largest singular value of a
M ×K submatrix ΦT , |T | = K, of an M ×N matrix Φ with i.i.d. subgaussian entries that are
properly normalized [16, 34, 35]:
P
(σmax(ΦT ) > 1 +
√K
M+ τ + β
)≤ e−Mτ2/2.
For large enough M , β ≪ 1; thus we ignore this small constant in the sequel. By letting
τ = jr√
1 + ǫK − 1−√
KM
(with the appropriate value of j for T ), we obtain
P(σmax(ΦT ) > jr
√1 + ǫK
)≤ e
−M2
“
jr√
1+ǫK−1−√
KM
”2
.
We use a union bound over all possible Rj supports for u ∈ Rj,K to obtain the probability that
Φ does not amplify the norm of u by more than jr√
1 + ǫK :
P(‖Φu‖2 >
(jr√
1 + ǫK)‖u‖2 ∀ u ∈ Rj,K
)≤ Rje
− 12(
√M(jr
√1+ǫK−1)−
√K)
2
.
Bound the right hand side by a constant µ; this requires
Rj ≤ e12(
√M(jr
√1+ǫK−1)−
√K)
2
µ (20)
for each j. We use another union bound among the residual subspaces Rj.K to measure the
probability that the RAmP does not hold:
P(‖Φu‖2 >
(jr√
1 + ǫK)‖u‖2 ∀ u ∈ Rj,K , ∀ j, 1 ≤ j ≤ ⌈N/K⌉
)≤⌈N
K
⌉µ.
To bound this probability by e−t, we need µ = KNe−t; plugging this into (20), we obtain
Rj ≤ e12(
√M(jr
√1+ǫK−1)−
√K)
2K
Ne−t
34
for each j. Simplifying, we obtain that for Φ to posess the RAmP with probability 1− e−t, the
following must hold for all j:
M ≥ 1(jr√
1 + ǫK − 1)2
(√
2
(lnRjN
K+ t
)+√K
)2
. (21)
Since (√a +√b)2 ≤ 2a + 2b for a, b > 0, then the hypothesis (12) implies (21), proving the
theorem.
APPENDIX II
PROOF OF THEOREM 5
In this proof, we denote M(x,K) = xK for brevity. To bound ‖Φ(x− xK)‖2, we write x as
x = xK +
⌈N/K⌉∑
j=2
xTj,
where
xTj= xjK − x(j−1)K , j = 2, . . . , ⌈N/K⌉
is the difference between the best jK model approximation and the best (j − 1)K model
approximation. Additionally, each piece xTj∈ Rj,K . Therefore, since Φ satisifes the (ǫK , s− 1)
RAmP, we obtain
‖Φ(x− xK)‖2 =
∥∥∥∥∥∥Φ
⌈N/K⌉∑
j=2
xTj
∥∥∥∥∥∥2
≤⌈N/K⌉∑
j=2
‖ΦxTj‖2 ≤
⌈N/K⌉∑
j=2
√1 + ǫKj
s−1‖xTj‖2. (22)
Since x ∈Ms, the norm of each piece can be bounded as